README.md

---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: detections
    list:
    - name: class
      dtype: string
    - name: bbox_xyxy
      list: float64
    - name: confidence
      dtype: float64
  - name: yolo_labels
    dtype: string
  - name: image_annotated
    dtype: image
  splits:
  - name: train
    num_bytes: 11360605942.32
    num_examples: 69018
  download_size: 11350785810
  dataset_size: 11360605942.32
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Forest Fire Detection Dataset — Auto-Annotated

Bounding-box annotated version of [touati-kamel/forest-fire-dataset](https://huggingface.co/datasets/touati-kamel/forest-fire-dataset),
built for training forest-fire / smoke / fog object detection models.

## Overview

This dataset contains video frames auto-labeled with bounding boxes for fire and
smoke-related visual phenomena, using a zero-shot open-vocabulary object detector
(Grounding DINO). It is derived from the original `touati-kamel/forest-fire-dataset` image
classification dataset, which did not include bounding box annotations.

## Classes

| Class          | Description                                   |
|----------------|------------------------------------------------|
| `Fire`         | Visible flame                                   |
| `Fire-smoke`   | Smoke originating from fire                     |
| `Fog`          | Fog / mist in the scene                         |
| `Factory-smoke`| Industrial/factory smoke (non-fire smoke source)|

## Why a single `train` split?

The source frames were extracted from videos and then shuffled randomly before
being split into train/validation/test. Because consecutive video frames are
often 99%+ visually similar, this shuffling caused near-duplicate frames from
the same video clip to end up scattered across different splits -- a data
leakage problem that would make validation/test metrics unreliable (a model
could "memorize" a near-identical frame seen during training).

To fix this, all annotated frames from the original train/validation/test
splits have been merged into a single `train` split here. **Validation and
test splits will be added later**, sourced from separate, distinct videos not
present in `train`, to ensure clean evaluation without leakage.

## Annotation methodology

- **Model**: `IDEA-Research/grounding-dino-tiny` (zero-shot, open-vocabulary
  object detection), run via Hugging Face `transformers`.
- **Prompts used** (mapped to class names):
  - `"flame"` → `Fire`
  - `"smoke from fire"` → `Fire-smoke`
  - `"fog"` → `Fog`
  - `"industrial smoke"` → `Factory-smoke`
- **Thresholds**: box confidence >= 0.30, text matching threshold >= 0.25.
- **Important**: these are automatically generated (teacher-model) annotations,
  **not human-verified**. Expect some false positives/negatives, especially on
  visually ambiguous frames (heavy haze, distant smoke, low light). Manual
  review or a secondary verification pass is recommended before using this
  data for anything beyond bootstrapping a first model.

## Schema

| Column     | Type          | Description                                                        |
|------------|---------------|----------------------------------------------------------------------|
| `image`    | `Image`       | Original, unannotated frame                                        |
| `detections` | `list[dict]` | One entry per detected box: `{"class": str, "bbox_xyxy": [x1,y1,x2,y2], "confidence": float}` |
| `yolo_labels` | `string`   | Same boxes pre-converted to YOLO format (`class_id x_center y_center width height`, normalized 0-1), one line per box, ready to write directly to `.txt` label files |
| `image_annotated` | `Image` (optional, some chunks) | Visual copy of `image` with boxes/labels drawn, for quick QA |

Class-to-ID mapping for `yolo_labels` is stored in `classes.json` at the repo root:
`{"0": "Fire", "1": "Fire-smoke", "2": "Fog", "3": "Factory-smoke"}` (order-dependent list).

## Usage

```python
from datasets import load_dataset

ds = load_dataset("touati-kamel/forest-fire-annotations")
example = ds["train"][0]
print(example["detections"])
print(example["yolo_labels"])
```

### Converting to a YOLO training folder

```python
import os

os.makedirs("yolo_dataset/images/train", exist_ok=True)
os.makedirs("yolo_dataset/labels/train", exist_ok=True)

for i, example in enumerate(ds["train"]):
    example["image"].save(f"yolo_dataset/images/train/{i:07d}.jpg")
    with open(f"yolo_dataset/labels/train/{i:07d}.txt", "w") as f:
        f.write(example["yolo_labels"])
```

## Roadmap

- Add genuinely separate `validation` and `test` splits from new, distinct
  video sources (not derived from frames already in `train`).
- Optional human-in-the-loop verification pass on a sample of auto-labeled
  boxes to estimate label quality/precision.

## Source data

Original unannotated frames: [touati-kamel/forest-fire-dataset](https://huggingface.co/datasets/touati-kamel/forest-fire-dataset)

## Maintainer

Kamel Touati ([HuggingFace: touati-kamel](https://huggingface.co/touati-kamel),
[GitHub: KamelTouati](https://github.com/KamelTouati))